New Zealand climate change information derived by multivariate statistical and artificial neural networks approaches

Climate change impact assessment requires data at the spatial and the temporal resolution at which impacts occur. The outputs of the current global climate models (GCMs) cannot be used directly in the development of desired climate change scenarios due to their coarse resolution. Artificial neural network (ANN) and multivariate statistics (MST) models were adapted to derive changes of site precipitation and temperature characteristics in a comparative study of their potential in downscaling GCM outputs. They were calibrated and verified using observed site temperature and precipitation over New Zealand and circulation variables from the National Center for Environment Prediction (NCEP) reanalysis. Subsequently, the models were used to derive changes of mean monthly precipitation and temperature characteristics from circulation variables projected in a transient climate change experiment performed by the Hadley Centre Global Climate Model (HadCM2). HadCM2 validated well with respect to NCEP reanalysis for its 'present climate' simulation. The predicted changes in seasonal mean sea level pressure fields over the 'New Zealand region' include intensified westerly flow particularly in winter. Therefore, the HadCM2 outputs contain information that can be used to derive localized climate change characteristics. Both downscaling models capture similar general patterns of change from the same GCM output, although they show substantial differences in localized characteristics, particularly for precipitation. The MST model suggests greater warming than the ANN model but both show larger temperature rises in summer and spring than in other seasons. Precipitation changes have marked spatial variation in their magnitude and direction; although by 2070-2099 more sites show precipitation increases than decreases, except for spring, and precipitation increases are more common in the North Island than the South Island. This intercomparison of two downscaling methods demonstrates an important source of uncertainty in regional climate change scenarios in addition to differing regional responses of GCMs. Copyright (C) 2001 Royal Meteorological Society.